Thematic Accuracy Assessment Procedures. Version 2 - USGS

to the distribution of sporadic or concentrated phenomena (Cochran 1977) and that can capture
some vegetation types that are underrepresented in the primary strata.
2.5 Determining Sample Sizes
2.5.1 Determining Acceptable Levels of Error and Confidence
Accuracy of a vegetation map may be evaluated from the perspective of whether the accuracy of
an entire map or the accuracies of its individual classes meet some specified threshold of
minimal accuracy, as in evaluating project performance. In such a case, expressing the results of
the assessment as a hypothesis test would be appropriate. Since the NPS Vegetation Inventory
accuracy assessment process is now more narrowly defined as a user information tool after the
project, rather than a project performance tool, this aspect of evaluating accuracy is now covered
by the validation step.
By the start of the thematic accuracy assessment campaign for a park, the map will have been
determined to have met a minimum accuracy threshold. The objective of the NPS Vegetation
Inventory thematic accuracy assessment is to inform the map user of the accuracy (limitations) of
individual map classes with a reasonably reliable estimate of the true accuracy of each map class.
The reliability of the accuracy estimate is reflected in its precision.
Precision is a measure of dispersion of the probability distribution associated with a
measurement. For the precision of an accuracy estimate (e.g., a mean from a sample) to be
understood, a confidence interval is required. A confidence interval is an interval within which
we have a specific level of confidence that the true value of an estimate lies; the narrower the
confidence interval associated with an estimate, the more precise is the estimate. The width of a
confidence interval is affected by (1) the sample size used to derive the point estimate, (2) the
confidence level that the user selects, and, (3) for binomial distributions, the value obtained for
the point estimate derived from the sample (see Figure 2). The first two factors may be
controlled by the investigator conducting the sampling; the third factor is a property of the
inference population being sampled and is often unknown or poorly known by the investigator
during the planning of the sampling. Larger sample sizes will result in a narrower confidence
interval, as will lower confidence levels. Smaller sample sizes and higher confidence levels will
widen the confidence interval. For binomial distributions, point estimates of the mean that
approach 0.5 (50%) will widen the confidence interval, whereas, with binomial (percentage)
data, estimates that approach either 0.0 (0%) or 1.0 (100%) will narrow the interval. Figure 2
shows the relationship between the sample size and confidence interval width for confidence
levels of 99%, 95%, and 90%, for two somewhat extreme values of pˆ : 0.5 (50%) and 0.9 (90%).
Of the parameters that can be controlled by the study design, the acceptance of a lower
confidence level can narrow the confidence interval width, as does increasing sample size
(particularly if the starting ample size is small). Confidence levels are often held fixed for a
given study (i.e., all values are reported to a predetermined level of confidence), where
conventionally used confidence levels are 90%, 95%, or 99%. The NPS Vegetation Inventory
uses a confidence level of 90%. Therefore, the width of the confidence interval will vary with
changing sample size. Since thematic accuracy sample size for a map class varies with class
abundance for the NPS Vegetation Inventory (Section 2.5.2), confidence intervals for map
classes that occupy fewer than 50 hectares will be wider than those for more abundant classes.
25